Unveiling Deep Learning Models: Leveraging Active Learning and High-Confidence Feature Extraction in Model Stealing

Recent research reveals that weaknesses in black box models can be exposed by stealing their functional characteristics. However, the current stealing process faces challenges such as low accuracy, insufficient hard label guidance and limited query quantity. To overcome these limitations, we have developed a method that facilitates the stealing of hard-label black-box models, called ALHC (Active Learning and High- confidence Feature Extraction). It incorporates a high-confidence stealing module to generate high-quality datasets. And it also introduces an efficient active learning selection strategy for training substitute models that exhibit higher similarity within the same query budget. In multiple datasets and real-world APIs, the ALHC has demonstrated remarkable effectiveness, surpassed traditional approaches and improved consistency and accuracy.